Should LLMs just treat text content as an image?